Introduction

Methods

TRIFID (Tool to Reliable Identification of Functional Isoform Data) is a Random Forest based predictor of the relative functional importance of splice isoforms.

It has been trained on reliable peptide evidence from 497 genes from the largest tissue-based proteomics analysis to date (79 experiments) and 47 features categorized in 5 groups (genome annotation, structural, splicing impact, cross-species conservation and RNA-seq expression).

Code repository available on github.com/fpozoc/trifid.

Results

SHAP feature importances and model interpretation

Figure 1: SHAP values (Lundberg et al. 2020) added by category. Features that best distinguish functional isoforms in the training set are conservation-based. Other important features include the length difference between the alternative isoform and the longest isoform, whether or not the transcript has a CCDS (Pruitt et al. 2009), and the conservation of Pfam functional domains. The SHAP scores can also provide clues to the influence of features on individual predictions (see example: Fibroblast growth factor receptor 1 (TRIFID Scores)

Functional importance in the human genome

Figure 2: Non-redundant isoforms divided into PRINCIPAL or ALTERNATIVE according to their annotation in APPRIS (Rodriguez et al. 2017). Most ALTERNATIVE isoforms have TRIFID scores below 0.05. Most PRINCIPAL isoforms have predictor scores above 0.9.

This work has been funded by the US National Institutes of Health grant 2 U41 HG007234.

Validating the model against an external source of information

Figure 3: Non-synonymous to synonymous (NS/Syn) ratios for alternative exons (exons that are exclusively present in alternative isoforms). Exons under selective pressure should have significantly lower NS/Syn ratios for common than for rare allele frequencies. The vast majority of alternative exons are not under selective pressure.

Conclusions

TRIFID discriminates functionally important isoforms with high confidence (MCC=0.89, AUPRC=0.98 over 5-folds CV of the training set). Thanks to TRIFID, we can now list the most biologically relevant alternative isoforms. Our model can be successfully exported to different genome species and genome annotation databases.

TRIFID predicts that a large majority of splice variants (85-90%) in the human genome are likely to not be functionally important at the protein level.

NS/Syn ratios show that exons from the highest scoring of alternative transcripts are under selective pressure, while low scoring exons have little or no evidence of selection.